feat(llm): allow setting query transformation behavior in BaseRAGQuestionAnswerer / BaseRAGQA (#67)

[omm-prakash18]

Summary

Fixes #67

This pull request introduces configurable query transformation behavior to BaseRAGQuestionAnswerer (and exposes the BaseRAGQA class alias).

Users can now configure whether and how user queries are rewritten or expanded before vector index retrieval (supporting standard query rewriting, HyDE - Hypothetical Document Embeddings, or custom prompt callables/UDFs), defaulting to None for backward compatibility.

---

What Changed

• Configurable query_transform:
  • Added query_transform: str | Callable | pw.UDF | None = None parameter to BaseRAGQuestionAnswerer.__init__ and AdaptiveRAGQuestionAnswerer.__init__.
  • Added robust resolution:
    • None (default): Skips transformation and queries the vector index with the raw user prompt.
    • "rewrite": Uses prompts.prompt_query_rewrite to generate keyword/entity-optimized search queries.
    • "hyde": Uses prompts.prompt_query_rewrite_hyde for Hypothetical Document Embeddings.
    • Callable / pw.UDF: Allows arbitrary custom transformation logic.
• Query Pipeline Execution:
  • In answer_query, when query_transform is active, the query is transformed with the LLM before calling self.indexer.retrieve_query. The raw user query is preserved for final answer generation.
• Ergonomic Alias:
  • Exported BaseRAGQA = BaseRAGQuestionAnswerer.
• Unit & Integration Tests:
  • Added test cases in python/pathway/xpacks/llm/tests/test_rag.py covering default None, "rewrite", "hyde", custom functions, invalid arguments, and the BaseRAGQA alias.

---

Example Usage

from pathway.xpacks.llm.question_answering import BaseRAGQA

# 1. Standard Query Rewriting
rag = BaseRAGQA(
    llm=chat_model,
    indexer=vector_server,
    query_transform="rewrite",
)

# 2. HyDE (Hypothetical Document Embeddings)
rag_hyde = BaseRAGQA(
    llm=chat_model,
    indexer=vector_server,
    query_transform="hyde",
)

# 3. Raw Query / Direct Search (Default)
rag_default = BaseRAGQA(
    llm=chat_model,
    indexer=vector_server,
    query_transform=None,
)

[CLAassistant]

All committers have signed the CLA.